Loading...
Loading...
Cursor and Moonshot have rolled out Composer 2.5, a substantial upgrade focused on long-horizon work, stronger instruction-following, and improved collaborative behavior. Built on the open-source Kimi K2.5 checkpoint, the release stems from scaled training, richer reinforcement-learning environments, targeted textual feedback, and synthetic data to refine localized behaviors and coding ability. Moonshot is also partnering with SpaceXAI to train a far larger model on Colossus 2 with roughly ten times the compute. Composer 2.5 introduces tiered pricing and a faster premium variant, highlighting practical strides toward more reliable, usable large models for sustained, complex tasks.
Composer 2.5 targets core pain points for deploying LLMs in production: sustained, long-horizon tasks and reliable instruction-following, which matter for automation, copilots, and multi-step workflows. Tech professionals should note improvements in collaboration and coding behaviors that can reduce human-in-the-loop overhead and integration risk.
Dossier last updated: 2026-05-18 18:02:15
Kimi, developer of the K2.5 AI model and related developer tools, is closing a $2 billion financing round that now includes multiple state-backed investors and central SOEs such as China Mobile and funds like Guozhitou and the Beijing AI Fund. The entry of national-capital players completes a restructuring of Kimi’s capitalization alongside prior strategic internet and industry investors. On the product side, Kimi’s Composer 2.5—built on the K2.5 model—was integrated into the popular coding app Cursor, signaling commercial deployment and developer adoption. The deal highlights state support for domestic AI platforms and could accelerate Kimi’s commercialization, partnerships, and competitive positioning in China’s AI ecosystem.
Elon Musk invited users to test Cursor’s new Composer 2.5 model, which Cursor calls its most powerful release and says was partially trained using Colossus 2 and built atop Kimi K2.5. Composer 2.5 focuses on long-task stability, complex instruction following, and collaboration, introducing text-feedback-guided RL that inserts short feedback at error points and uses distilled KL loss to correct local mistakes. Cursor scaled synthesis training 25× versus Composer 2 to boost coding ability and uses held-out-function tests as reward signals, while warning of reward-cheating risks. Training infrastructure includes sharded Muon, dual-grid HSDP, asynchronous all-to-all to overlap comms and compute, and mixed HSDP layouts; pricing is per-million-token input/output with standard and fast tiers.
Mouse Labs' Composer 2.5, deployed in Cursor, is a major update improving sustained work on long tasks, following complex instructions, and collaborative behavior versus Composer 2. The team scaled training, expanded RL environments, and added new learning methods including targeted textual feedback—injecting hints into local context to create a teacher distribution and applying a KL distillation loss to correct specific mistakes during long rollouts. Composer 2.5 also uses 25× more synthetic coding tasks and harder, dynamically generated problems to push capability. Mouse Labs and SpaceXAI are jointly training a much larger model (Colossus 2 hardware with a million H100-equivalents) using 10× more compute, aiming for a significant capability leap.
Cursor Introduces Composer 2.5
Cursor Introduces Composer 2.5
Cursor : Cursor releases Composer 2.5, saying it's better at sustained work on long-running tasks and follows complex instructions more reliably; it's built on Kimi K2.5 — Composer 2.5 is now available in Cursor. — It's a substantial improvement in intelligence and behavior over Composer 2.
Composer 2.5, a major update from Moonshot available in Cursor, improves sustained task performance, instruction-following, and collaborative behavior versus Composer 2. The team scaled training, created more complex RL environments, and added new learning methods including targeted textual feedback and synthetic data to refine localized behaviors and coding ability. Composer 2.5 uses the same open-source checkpoint (Kimi K2.5) and Moonshot is partnering with SpaceXAI to train a much larger model from scratch on Colossus 2 with roughly 10x more compute. Pricing tiers are announced ($0.50/$2.50 per M input/output tokens; a faster, costlier variant is also offered). These advances matter for real-world usability of large models and more reliable long-horizon decision-making.